1. English-Hindi Neural machine translation and parallel corpus generation EKANSH GUPTA ROHIT GUPTA

2. Advantages of Neural Machine Translation Models Require only a fraction of the memory needed by traditional statistical machine translation (SMT) models Deep Neural Nets out-perform previous state of the art methods assuming availability of large parallel corpora Can be combined with word-alignment approach to address the rare-word problem

3. Performance of NMT Models Source: T Luong et al, ACL 2015

4. Motivation Advantages of Neural Machine Translation Very large parallel English-Hindi corpora are unavailable However comparable corpora available

5. Encoding Scheme Using one-hot encoding for top-N words chosen from large monolingual corpora for each language Out Of Vocabulary (OOV) words represented by unk Monolingual corpora used: http://corpora.heliohost.org/

6. Encoding Scheme: Example (top 10 words) the 1000000000…... to 0100000000…... and 0010000000…... a 0001000000…... of 0000100000…... in 0000010000…... for 0000001000…... that 0000000100…... is 0000000010…... on 0000000001…... के 1000000000…... में 0100000000…... की 0010000000…... को 0001000000…... से 0000100000…... है 0000010000…... ने 0000001000…... का 0000000100…... और 0000000010…... कि 0000000001…...

7. Recurrent Neural Networks

8. Summary of Methodology Training weak translator using limited parallel corpus Weak translator and aligning heuristic (ex: Hunalign) used to create additional parallel corpus Neural translator re-trained on generated bigger parallel corpus

9. Pipeline for creating parallel corpora Source: K Wołk et al, 2014

10. Aligning sentences Original Hindi Translated English Original English Comparison Heuristics Basic Translation Engine Comparison Adapted From: K Wołk et al, 2014

11. Questions ?